Hiptmair–Xu Preconditioners

• The paper presents Hiptmair–Xu preconditioners as auxiliary-space methods that decompose complex H(curl) and H(div) operators into simpler nodal and smoothing components.
• They replace difficult edge- or face-element solves with additive combinations of Jacobi smoothers and nodal H¹-type corrections, ensuring coercivity or proper treatment of singular kernels.
• The approach extends to mixed discretizations, interface substructuring, and nonstandard complexes, providing scalable and robust solvers across various Maxwell and saddle-point problems.

Hiptmair--Xu preconditioners are nodal auxiliary-space preconditioners for discrete $H(\curl)$ and $H(\div)$ operators. Their defining idea is to replace a difficult edge- or face-element solve by an additive combination of a simple smoother on the original space, one or more scalar or vector $H^1$-type auxiliary solves in nodal spaces, and exact-sequence corrections through discrete gradient or curl maps. In the positive Maxwell regime this produces coercive preconditioned operators, while in semidefinite curl-curl problems the same architecture resolves the large gradient kernel through auxiliary-space splitting rather than direct inversion of the singular operator (Delville-Atchekzai et al., 2023, Hu, 2017).

1. Core auxiliary-space architecture

In the positive Maxwell setting on a bounded polyhedral domain $\Omega\subset\mathbb R^3$, the discrete operator has the curl-curl-plus-mass form

$$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$

with $W(\Omega)$ the lowest-order Nédélec edge element space. The added positive mass term makes the operator coercive on $W(\Omega)$, which is precisely the regime in which HX is classically formulated as a positive Maxwell preconditioner (Delville-Atchekzai et al., 2023).

A standard HX formula recalled in later work is

$$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$

Here $\operatorname{diag}(\mathscr M)^{-1}$ is a Jacobi smoother on the edge space, $G_\Omega:V(\Omega)\to W(\Omega)$ is the discrete gradient from the nodal $H(\div)$0 space, and each $H(\div)$1 is a nodal-to-edge transfer induced by edge interpolation. This realizes the characteristic HX decomposition into one edge-space smoothing term, one gradient correction, and three nodal lifting terms (Delville-Atchekzai et al., 2023).

In the weighted Maxwell formulation with jump coefficients, the same three-part pattern is written as

$H(\div)$2

The factors are the Jacobi smoother $H(\div)$3, a vector nodal Laplacian solve transferred by the edge interpolation $H(\div)$4, and a gradient-space correction through the restriction of $H(\div)$5 to $H(\div)$6. The paper emphasizing jump coefficients states that implementation requires solving four scalar elliptic problems, which is one of the practical attractions of the method (Hu, 2017).

A closely related formula appears in recent time-harmonic practice for a nearby positive Maxwell block: $H(\div)$7 This variant is not a new HX theory; it is an implementation of the same auxiliary-space principle in a real-valued split formulation, with a smoother, a nodal/vector-Laplacian correction, and a scalar-gradient correction (Fressart et al., 17 Jul 2025).

2. Regular decomposition, exact sequences, and robustness

The analytical foundation of HX is a discrete regular decomposition. In the edge-element setting one seeks, for $H(\div)$8, a splitting of the form

$H(\div)$9

with $H^1$0 in a nodal space, $H^1$1 in a vector nodal space, and $H^1$2 a remainder controlled by smoothing. Later auxiliary-space analyses identify this decomposition as the exact mechanism matching the three blocks of the preconditioner (Hu, 2017, Hu, 2017).

For heterogeneous Maxwell problems with strongly discontinuous coefficients, the relevant norms are coefficient-weighted: $H^1$3 The key difficulty is that classical unweighted decompositions do not control the effect of large jumps in $H^1$4 and $H^1$5. The weighted theory therefore introduces upper intersection sets

$H^1$6

together with the geometric notion of thorny vertices, which are the obstruction to full-space weighted stability (Hu, 2017).

The no-thorny case yields the sharpest HX results. If the coefficients belong to the class $H^1$7, every $H^1$8 admits

$H^1$9

with

$\Omega\subset\mathbb R^3$0

and

$\Omega\subset\mathbb R^3$1

This leads to

$\Omega\subset\mathbb R^3$2

and in the best geometric case,

$\Omega\subset\mathbb R^3$3

The hidden constants depend on geometry, mesh regularity, and the class parameter $\Omega\subset\mathbb R^3$4, but not on the jump magnitudes themselves (Hu, 2017).

The companion decomposition paper develops the local ingredients behind these estimates. It proves regular Helmholtz-type decompositions for lowest-order Nédélec functions on ordinary polyhedra, on unions of polyhedra meeting at an edge or at a vertex, and under constraints prescribing zero tangential degrees of freedom on selected faces and edges. The characteristic estimates are of the form

$\Omega\subset\mathbb R^3$5

with the logarithmic factor removable in several favorable cases. This provides the interface-compatible splitting toolkit used later in the weighted HX convergence theory (Hu, 2017).

3. Semidefinite curl-curl operators and saddle-point systems

HX is not restricted to strictly positive operators. An auxiliary-space theory for semidefinite linear systems formulates the preconditioner abstractly as

$\Omega\subset\mathbb R^3$6

with $\Omega\subset\mathbb R^3$7 surjective and $\Omega\subset\mathbb R^3$8 simpler than $\Omega\subset\mathbb R^3$9. For semi-SPD $$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$0, the theory replaces ordinary condition-number formulas by range-restricted identities involving an infimum over $$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$1. This makes the kernel treatment explicit rather than implicit (Park et al., 8 Sep 2025).

In that framework, the $$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$2 HX preconditioner for the SPD case $$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$3 is written as

$$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$4

while in the semidefinite case $$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$5 the gradient correction is dropped: $$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$6 The paper proves

$$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$7

and, analogously for $$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$8,

$$\langle \mathscr M(u),v\rangle = \int_\Omega \boldsymbol{curl}(u)\cdot \boldsymbol{curl}(v)+\gamma^2 u\cdot v\,dx, \qquad u,v\in W(\Omega),$$9

thereby recovering uniform optimality in both SPD and semidefinite regimes through the same auxiliary-space variational identities (Park et al., 8 Sep 2025).

A distinct semidefinite application arises in the mixed finite element discretization of time-harmonic Maxwell saddle-point systems at vanishing wave number. For

$W(\Omega)$0

with $W(\Omega)$1 the singular discrete curl-curl matrix, $W(\Omega)$2 spanning $W(\Omega)$3, and $W(\Omega)$4, the exact inverse formula yields

$W(\Omega)$5

and reduces the saddle-point problem to

$W(\Omega)$6

The crucial point is that the right-hand side lies in $W(\Omega)$7, so the singular curl-curl equation is consistent. The paper then states that one may compute an arbitrary particular solution $W(\Omega)$8 by CG with the Hiptmair-Xu preconditioner and recover the full mixed solution by explicit kernel corrections. This is not a new HX construction for the full saddle-point matrix; it is an exact algebraic reduction showing when HX is the natural solver for the singular curl-curl block (Xiang et al., 2016).

4. Substructured and interface HX preconditioners

Classical HX is a volume preconditioner, but it can be transferred to a nonoverlapping substructuring setting. For positive Maxwell problems, a central exact identity for the scalar Schur complement is

$W(\Omega)$9

where $W(\Omega)$0 is the block-diagonal local Schur inverse and $W(\Omega)$1 is the interface assembly operator. This identity says that tracing the inverse of the global conforming scalar volume operator to the skeleton gives exactly the inverse of the assembled scalar Schur complement, not merely a spectrally equivalent approximation (Delville-Atchekzai et al., 2023).

Once this identity is available, the volume HX decomposition can be pushed term-by-term to the interface. The resulting skeleton analogue is

$W(\Omega)$2

Replacing the exact scalar Schur inverse by a Neumann–Neumann preconditioner

$W(\Omega)$3

gives the practical substructured HX preconditioner

$W(\Omega)$4

The architecture is recognizably HX, but every nodal inverse has been replaced by a scalar interface Schur-complement preconditioner (Delville-Atchekzai et al., 2023).

The corresponding estimate is

$W(\Omega)$5

This shows that the interface Maxwell preconditioner inherits the HX quality factor from the volume problem and multiplies it by the scalar Schur-complement quality factor. The same paper explicitly notes that because the scalar ingredient is only one-level Neumann–Neumann, the method is not expected to be fully scalable in the strong sense; adding a coarse space is identified as the natural improvement (Delville-Atchekzai et al., 2023).

5. Generalizations to nonstandard complexes and meshes

The HX paradigm has been extended well beyond lowest-order simplicial volume discretizations. For lowest-order mixed virtual element methods on polytopal meshes, the discrete complex is exact,

$W(\Omega)$6

the cochain projection commutes,

$W(\Omega)$7

and one has a VEM regular decomposition

$W(\Omega)$8

with

$W(\Omega)$9

This leads, in the 3D edge case, to the direct HX analogue

$$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$0

and to an $$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$1 facet preconditioner with a recursive $$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$2 structure. The spectral condition number is proved to be bounded independently of $$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$3, while robustness with respect to high aspect ratio is reported as a numerical observation rather than a theorem (Boon et al., 2024).

On triangulated surfaces, the geometry changes the auxiliary space itself. A central observation is that any vector field tangential to a triangulated surface $$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$4 cannot be continuous. The surface HX construction therefore uses the ambient nodal space $$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$5 together with a modified interpolation $$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$6 for the $$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$7 branch. The additive surface preconditioner is

$$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$8

and satisfies

$$\tilde{\mathscr M}^{-1} = \operatorname{diag}(\mathscr M)^{-1} + G_\Omega \mathscr L^{-1} G_\Omega^* + \sum_{j=1,2,3} \Pi_\Omega^{e_j}\mathscr L^{-1}(\Pi_\Omega^{e_j})^* .$$9

For 2D surfaces the method uses four scalar surface Laplacian inverses: three from the ambient nodal-vector block and one from the exact-sequence term (Li, 2021).

A further extension replaces the fixed-dimensional de Rham complex by a mixed-dimensional one on networks of manifolds. There the mixed-dimensional regular decomposition takes the form

$\operatorname{diag}(\mathscr M)^{-1}$0

with

$\operatorname{diag}(\mathscr M)^{-1}$1

The corresponding mixed-dimensional HX preconditioner is

$\operatorname{diag}(\mathscr M)^{-1}$2

with

$\operatorname{diag}(\mathscr M)^{-1}$3

This suggests that the HX mechanism is fundamentally a de Rham-complex construction rather than a feature specific to standard 3D $\operatorname{diag}(\mathscr M)^{-1}$4 discretizations (Budisa et al., 2019).

6. Practical use, indefinite Maxwell problems, and abstract operator-preconditioning

In large-scale time-harmonic Maxwell computations, HX is often used indirectly rather than as a preconditioner for the original indefinite and non-Hermitian operator. In one recent parallel study, the complex system $\operatorname{diag}(\mathscr M)^{-1}$5 is rewritten in split real-imaginary form and preconditioned by the block-diagonal positive surrogate

$\operatorname{diag}(\mathscr M)^{-1}$6

Each positive block is then approximately inverted by hypre’s AMS, which the paper identifies as HX in that software stack. The outer solver is right-preconditioned FGMRES because the inner block inverses are computed inexactly (Fressart et al., 17 Jul 2025).

That study reports two HX variants. In the first, each positive block solve uses CG preconditioned by HX with maximum $\operatorname{diag}(\mathscr M)^{-1}$7 inner iterations and relative residual $\operatorname{diag}(\mathscr M)^{-1}$8. In the second, AMS is used directly as the approximate inner solver. For $\operatorname{diag}(\mathscr M)^{-1}$9, the outer FGMRES iteration counts for the inner-CG variant are $G_\Omega:V(\Omega)\to W(\Omega)$0 across refinements from $G_\Omega:V(\Omega)\to W(\Omega)$1 to $G_\Omega:V(\Omega)\to W(\Omega)$2 unknowns, which the paper interprets as essentially mesh-independent. For $G_\Omega:V(\Omega)\to W(\Omega)$3, the outer counts remain roughly mesh-independent, around $G_\Omega:V(\Omega)\to W(\Omega)$4–$G_\Omega:V(\Omega)\to W(\Omega)$5, but the method deteriorates with increasing wavenumber and even more strongly with increasing physical domain size. The authors state that the number of FGMRES iterations is at least proportional to $G_\Omega:V(\Omega)\to W(\Omega)$6, and on an enlarged-domain test the iteration count is multiplied by $G_\Omega:V(\Omega)\to W(\Omega)$7 relative to the small-domain case (Fressart et al., 17 Jul 2025).

These observations clarify a common misconception. HX is tailored to positive Maxwell equations, not to the original indefinite time-harmonic scattering operator. In that setting it is therefore being used as inner technology for a nearby positive block problem, not as a direct preconditioner for $G_\Omega:V(\Omega)\to W(\Omega)$8. The same paper nevertheless finds HX and Block Low-Rank preconditioners to be the most promising iterative approaches among those tested, while also showing that BLR can be substantially faster on a large-domain benchmark (Fressart et al., 17 Jul 2025).

At a more abstract level, operator-preconditioning theory explains why HX often tolerates inexact auxiliary solves. A Petrov–Galerkin extension of operator preconditioning considers a perturbed main operator $G_\Omega:V(\Omega)\to W(\Omega)$9 and a perturbed preconditioner $H(\div)$00, with the key estimate

$H(\div)$01

The main implication is that discretization accuracy and preconditioning accuracy can be decoupled: one may require $H(\div)$02 small for PDE accuracy while allowing $H(\div)$03 for a cheap or compressed auxiliary preconditioner. This is not an HX construction theorem, but it is directly relevant to practical HX implementations that replace exact nodal inverses by AMG cycles, incomplete factorizations, or other approximations (Escapil-Inchauspé et al., 2020).

Taken together, these developments portray Hiptmair--Xu preconditioners as a stable auxiliary-space paradigm whose essential ingredients are exact-sequence structure, a regular decomposition, and nodal $H(\div)$04-type auxiliary solvers. The same pattern appears in positive Maxwell problems, semidefinite curl-curl systems, saddle-point reductions, Schur-complement substructuring, virtual elements, surface complexes, and mixed-dimensional geometries. What changes from one setting to another is not the core mechanism, but the decomposition theory and transfer operators needed to make that mechanism stable and robust (Delville-Atchekzai et al., 2023, Park et al., 8 Sep 2025).